← All articles

AI Agent or Automation? The One Question That Tells You Which You Need

October 7, 2026 · 3 min read

Both get sold as "set it and forget it," and both can save you hours. But an automation and an AI agent are different tools, and picking the wrong one is how people end up with either a brittle mess or an expensive, unpredictable one. Here's the plain-English difference, and a quick way to choose.

An automation follows a recipe

A classic automation, the kind you build in Zapier, Make, or an email rule, is a fixed sequence: when X happens, do Y, then Z. A new form submission arrives, a row is added to a spreadsheet, a Slack message is posted. You decide every step in advance, and the tool executes them exactly the same way every time.

Anthropic's engineering team draws the same line when it describes AI systems. It calls the first kind workflows: "systems where LLMs and tools are orchestrated through predefined code paths." Notice that an AI model can still be inside a workflow, for example to summarize an email or sort it into a category. The path through the steps, though, is set by you, not by the model.

An agent chooses its own steps

An agent is the other kind. Anthropic defines agents as "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." In practice, you hand it a goal, such as "find out why last week's signups dropped," and it decides what to look at, which tools to use, and when it has enough to answer. If the first approach fails, it tries another.

That flexibility is the whole appeal. It's also the cost: the agent may take a different route each time, it can take more steps than you expected, and it can be wrong in ways a fixed recipe can't. Judgment is powerful, and judgment can misfire.

Which one to use: ask "can I write the steps down?"

Here's the test. Try writing the task as a numbered list you could hand to a new intern with no discretion. If you can, and the list rarely changes, use an automation. Examples include sending a receipt when a payment clears, copying new leads into your CRM, or posting a weekly reminder. It will be faster, cheaper, and easier to debug, because you can see exactly what happened at every step.

If the honest answer is "it depends on what they find," you're in agent territory. Researching a question across several sources, triaging messy requests that don't fit neat categories, or investigating a problem whose cause you don't know yet are all jobs where the steps can't be known ahead of time.

Many good setups combine the two. A fixed workflow handles the predictable skeleton (trigger, fetch, deliver), while a model or agent handles the one fuzzy step in the middle. Anthropic's advice is worth borrowing: find "the simplest solution possible, and only [increase] complexity when needed." Start with a recipe, and add judgment only where the recipe keeps breaking.

A practical way to start

Pick one repetitive task from your week and write out its steps. Highlight any step where you'd say "well, it depends." If nothing is highlighted, build a plain automation. If one step is highlighted, let an AI handle just that step inside the automation. If most steps are highlighted, try an agent, give it a narrow goal, limit what it can touch, and review its work until you trust it.

The takeaway: an automation is a recipe, an agent is a cook. Use the recipe when you know what the dish is, and the cook when you don't.

Sources: Building effective agents (Anthropic)